updating our beliefs about inconsistency: the monty-hall case · updating our beliefs about...
Post on 20-Jul-2020
1 Views
Preview:
TRANSCRIPT
HAL Id: ijn_00264949https://jeannicod.ccsd.cnrs.fr/ijn_00264949
Preprint submitted on 18 Mar 2008
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Updating our beliefs about inconsistency: TheMonty-Hall case
Jean Baratgin
To cite this version:
Jean Baratgin. Updating our beliefs about inconsistency: The Monty-Hall case. 2008. �ijn_00264949�
1
Updating our beliefs about inconsistency: The Monty-Hall case
1 Introduction
Since the sixties, various researchers have come to the conclusion that participants in
psychological studies revise their probability judgments differently from the way
experimenters do using Bayes' rule. People seem unable to revise their degrees of belief in a
Bayesian manner (for a review see Gilovich et al. 2002). They are thus thought to be
inconsistent. However this conclusion appears debatable when one sheds light on the notion
of Bayesian consistency. In such light, the methodology underlying the above-mentioned
studies does not seem concerned by individual consistency. It rather amounts to an analysis of
accuracy. The experimenter plays the role of an expert who forms ”correct predictions”
(Baratgin & Politzer 2006). The Monty-Hall Puzzle constitutes a much-studied paradigm
emblematic of this literature. The standard version originates from a popular television game
show “Let's Make A Deal” programmed in 1963 in North America. It involves only three
numbered doors and equal prior probabilities (from now on we will refer to it as standard
MHP)1. This paper will focus on a generalized probabilistic version of the Monty-Hall Puzzle
(that we will note MHPnn-2). It corresponds to a four stages story that reads as follows2:
Stage (i) A TV host shows you n numbered doors No.1, ...No.n, one hiding a car and the
other n-1 hiding goats. Let carNo.i be the event “the car is behind door No.i”. Let
1 It was phrased as a problem of choice where the host after picking door No.3 to reveal a goat, asksparticipant: “should you switch your original choice No.1 to No.2?”. Despite a choice is required inthis version, participants must first of all make a probabilistic judgment. An older version called theThree Prisoners Problem is cast in a probabilistic format (Gardner 1959). This paper focuses on themere probabilistic judgment disregarding the further question whether to switch. We will indifferentlymake references to the Three Prisoners Problem or to the Monty Hall Puzzle under the same notation.The question of choosing whether to switch door can be treated independently from the probabilityjudgment question. Chun (2003) for example analyses standard MHP as a decision problem.Conversely numerous experimental studies focus on the sole question of choice whether to switchdoor (see for example Gilovich et al. 1995; Granberg & Brown 1995; Friedman 1998).2 This version is an adaptation of standard MHP (see for example Gillman 1992).
2
€
P(carNo.i) =aiA
, i = 1, ..., i = n, with
€
A = ass=1
s=n∑ be the corresponding prior probabilities given
at this stage. The version with equal prior probabilities will be noted standard MHPnn-2.
Stage (ii) You get to pick a door, winning whatever is behind it. You choose door No.1,
say. What is the probability that the car is behind door No.1?
Stage (iii) The host, who knows where the car is, tells you “I will show you n-2 doors (out of
the n-1 other doors) that hide goats”.
Stage (iv) Then the host opens doors No.3, ... to No.n to reveal goats (a message that we
note ‘goatNo.3...n’ with single quotation marks), and asks you what is now the probability that
the car is behind door No.1?
Solving MHPnn-2 requires comparing the probabilities that the car is respectively
behind doors No.1 or No.2 given that, the host (who knows where the car is) has chosen to
open doors No.3... to No.n (out of the n-1 other possibilities) revealing goats behind.
In the various experiments on the Monty Hall Puzzle, participants’ modal probabilistic
response is
€
a1a1 + a2
(noted from now on MR). It departs from experimenters’ “Bayesian
solution” called “EBS”, which general form in MHPnn-2 establishes to
€
a1a1 + (n−1)a2
(see for
a review of studies on standard versions (with equal prior probabilities) Krauss & Wang
(2003); and for experiments that propose a version with unequal prior probabilities see:
Ichikawa (1989); Ichikawa & Takeichi (1990); Johnson-Laird, Legrenzi, Girotto & Sonino-
Legrenzi (1999); Granberg (1996); Yamagishi (2003)).
Most psychologists, economists and lawyers consider these studies as mounting
evidence of the inconsistency of human beings (Piattelli-Palmerini 1994; Friedman 1998;
Caplan 2000; Risinger & Loop 2002; Kluger & Wyat 2004). The purpose of this paper is to
3
analyze the rationale for such a conclusion. The aim here is not to conduct a thorough review
of the literature testing the experimental robustness of MR but rather to discuss the underlying
idea of participants’ inconsistency. In order to validate this often drawn conclusion,
experimenters should make sure that participants share their representation of MHPnn-2.
Alternatively, the response usually made by participants may be seen as consistent with
respect to a different representation of the puzzle3. In this view, specific normative4 solutions
can be determined for each particular representation (experimenters’ and participants’
ones).
This paper is organized in three sections. First, experimenters’ representation of
MHPnn-2 is presented. “EBS” is the correct solution in a traditional situation of probability
revision where the message ‘goatNo.3...n’ focuses attention on a given subset of the original set
of hypotheses. Second we explain that MR corresponds to a correct solution in an updating
interpretation of the situation of revision where the message modifies the structure of the
original problem. We show by paying attention to MHPnn-2’s statement that pragmatics
reasons can explain the biased interpretation of the revision message. In a third part, we show
that the different explanations proposed in the literature to descriptively account for
participants’ responses can be derived from an updating framework.
2 Experimenters’ representation of MHPnn-2
3 The hypothesis that participants’ response could be consistent with a representation of theexperimental paradigm different from what experimenters expect is rarely envisaged in the literatureon probabilistic revision. However a significant time ago various authors in the field of formal logicretained this idea to study the possible interpretations of the conditional. Indeed when participantsinterpret a conditional as a biconditional (If p then q leads one to infer that If not-p then not-q), theycorrectly proceed to an Affirmation of the Consequent although this affirmation would be consideredas an error by experimenters (Geis & Zwicky 1971; Ducrot 1972; Fillenbaum 1976).4 The term “normative” is taken here in the sense of a theoretical model considered as a referentialnorm for “rational” judgment. This is the case, for example, of the Subjective Expected Utility Theoryin the framework of decision theory, or of the Bayesian model in the context of probabilisticjudgment. The question asked within this methodology is whether man is “consistent” in hisjudgments and decisions with respect to the norm.
4
2.1 Basics postulates
Any empirical study is based on experimenters’ belief that participants accept with
certainty the data provided in the statement of the problem. In MHPnn-2, participants are
supposed to distinguish the four stages (i), (ii), (iii) and (iii) and to accept the explicit and
implicit rules related to these stages. The explicit rules are5:
Rule 1: The host has a distribution of preference (
€
P(carNo.i) =aiA
, with
€
A = ass=1
s=n∑ ) when he
initially places the car behind one door.
Rule 2: The host cannot open the door chosen by participants (and conventionally chosen as
No.1).
Rule 3: The host never opens a door (among the n-2 doors he or she will open) that hides the
car. This rule implies:
Rule 3.1: If the car is behind door No.i (with i = 2, ..., i = n), the host is bound to show
goats behind the n-2 doors No.2, ...,No.i-1, No.i+1, ..., No.n.
Rule 3.2: if the car is behind door No.1, the host is bound to show goats behind n-2
doors chosen among the n-1 doors No.2..., No.n. In accordance with the experimental
literature, we consider the additional following rule:
Rule 3.2.1: In the specific case where the car is behind door No.1, to show
goats, the host has no preference among the n-1 doors No.2, ..., No.n to choose n-2 doors.
Hence experimenters in MHPnn-2 form the following first postulate.
5 These rules are immediate translations of the rules applied to MHP and standard MHPnn-2 by allexperimenters. In Appendix 1 we discuss the difficulty of transposing the rule 3.2 in a general MHPnk(untreated by psychologists) where k doors with goats behind are opened (k ≤ n-2).
5
Postulate 1. Participants consider the four stages and accept the implicit and explicit rules 1,
2, 3, 3.1 and 3.2 implied by MHPnn-2’s statement.
The second set of experimenters’ beliefs, largely accepted in the great majority of
experimental paradigms on probabilities revision (see Baratgin & Politzer 2006), seems a
natural consequence of Postulate 1. Experimenters think that participants adhere to their two-
fold representation of MHPnn-2. To the mental construction of the various stages of MHPnn-2
(fixed system) based on the three first stages on the one hand and to the correct
comprehension of the message of revision given in stage (iv) on the other hand.
Postulate 2. Participants share experimenters’ representation of MHPnn-2 derived from by
postulate 1. They specifically have the same structure of believe (representation of system)
and the same interpretation of the message of revision.
Let’s next analyze experimenters’ double representation.
2.2 A bi-probabilistic level
The semantics of possible worlds as used in Economics and Artificial Intelligence
helps to model the belief structure experimenters expect from participants surrounded by
uncertainty (Walliser & Zwirn 1997; Billot & Walliser 1999; Fagin et al. 1999; Walliser &
Zwirn forthcoming). Due to uncertainty, participants envisage a plurality of worlds derived
from the alternative situations proposed by the experimental paradigm that they consider as
possibly real. In most experiments on probability revision, participants are requested to
form several layers of uncertain beliefs related to the various properties and meta-properties
characterizing the object of prediction (for a review, see Baratgin & Politzer 2007a)6. After
6 For instance, in the well-known “medical diagnosis” problem (Hammerton 1973), participants (castas physicians) are requested to form a belief about a patient. They consider the different symptoms(mammography positive, negative), the possibly related diseases (breast cancer or not) and the beliefwhether the patient suffers from the disease. Various other paradigms are designed in the same way,
6
the original participant’s choice of door No.1, MHPnn-2 can be modeled as a “two-level
hierarchy belief structure” (Walliser & Zwirn 1997, forthcoming).
For experimenters the participant in MHPnn-2 has to form uncertain beliefs on the n-2
doors to be opened by the host which depend on his (or her) uncertain beliefs on the door that
he (or she) believes to hide the car. In experimenters’ view, participants after to have picked
the door No.1 construct three layers of beliefs:
- A layer 0 illustrates the basic properties of the door (whether to be opened by the
host). It comprises the n-1 possible combinations offered to the host at stage (iii) when
choosing n-2 doors out of n-1. It corresponds to the following n-1 worlds {’goatNo.2.4…n’,
‘goatNo.2.3.5.6…n’, ..., ‘goatNo.2…n-1’, ‘goatNo.3…n’},
- A layer 1 accounts for the secondary properties of the door (whether to hide the car).
It comprises the n possibilities offered to the participant at stage (ii). It corresponds to the
following n worlds {carNo.1, ..., carNo.n},
- A layer 2 defines the door of prediction (the door originally picked by the
participant). It corresponds to the world {No.1}.
By Postulate 1, participants know that the host’s choice to open n-2 doors that hide
goats at stage (iii) depends on their choice at stage (ii). Hence, participants define a causal
relation between the two layers 0 and 1. Similarly, participants know that the host’s choice to
initially place the car is limited to n doors. Hence, participants define a populational relation
between layers 1 and 2.
Experimenters assume that participants translate these relations into two objective
distributions of probability elaborated from the explicit and implicit rules of MHPnn-2. The
see for example the “cab problem” (Kahneman & Tversky 1972), and the “engineer-lawyer problem”(Kahneman & Tversky 1973).
7
distribution of probability from layer 2 to 1 is explicitly defined by rule 1. It is represented by
the prior probabilities P(carNo.i) (with i = 1, ..., i = n). The distribution of probability (called
likelihoods) of layer 1 to 0 is precisely defined by rules 2, 3, 3.1 and 3.2. If the car is behind
door No.1, the host uses an indifference strategy and the n-1 worlds of layer 0 have the same
probability of 1/(n-1) of being opened to show a goat. If the car is behind door No.i (with i >
1), the host can open only the n-2 doors No.3...i-1.i+1...n.
We thus establish the first property of experimenters’ representation of MHPnn-2.
Property 1. Experimenters’ structure of belief of MHPnn-2 is a bi-probabilistic structure (see
Figure 1a)7 where the distributions of probabilities are consistent with rules 1, 2, 3, 3.1 and
3.2.
-------------------------------------------------------------
Insert Figure 1 about here
-------------------------------------------------------------
2.3 A focusing situation of revision
We can note that the bi-probabilistic structure of MHPnn-2 can be collapsed into an
equivalent probabilistic structure of one level with the 2(n-1) possible combined worlds (see
Figure 1b): {carNo.1∧’goatNo.2.4…n’, carNo.1∧’goatNo.2.3.5.6…n’, ..., carNo.1∧’goatNo.2…n-1’,
carNo.1∧’goatNo.3…n’, carNo.2∧’goatNo.3…n’, carNo.3∧’goatNo.2.4…n’, ..., carNo.n∧’goatNo.2…n-
1’}. Hence we have the following relations between the two structures: (1)
P(carNo.1) = P(carNo.1∧’goatNo.2.4…n’) + P(carNo.1∧’goatNo.2.3.5…n’) + ...
+ P(carNo.1∧‘goatNo.2…n-1’) + P(carNo.1∧‘goatNo.3…n’)
7 As in Walliser and Zwirn (forthcoming) a physical world is represented here by a square and apsychical world is represented by a circle. Here experimenters have a subjective belief thatparticipants’ beliefs match this objective structure and in level 2, the 2-world is represented by acircle.
8
P(carNo.2) = P(carNo.2∧ ‘goatNo.3…n’)
...
P(carNo.n) = P(carNo.n∧‘goatNo.2…n-1’).
In the last stage (iv) of MHPnn-2, participants learn that the host has opened doors No.3,
... to No.n showing goats behind following the protocol described in stage (iii) (among the n-1
possible doors No.2, ...and No.n). In experimenters’ opinion, this message indicates that the
subsystem where the world ’goatNo.3...n’ is true should be extracted from the initial system
(bi-probabilistic structure of Figure 1). Participants should form their beliefs about this
extracted subsystem according to message ‘goatNo.3…n’ by keeping the same distribution of
probabilities P. They are expected to focus their attention on such subset. Practically, this
subset comprises the two worlds {carNo.1∧’goatNo.3…n’, carNo.2∧’goatNo.3…n’}, the other n-2
following worlds being temporally left out. Thus participants should arrive to a one level
belief hierarchy corresponding to a simpler probabilistic structure where the basic worlds are
combined over layers 0 and 1 of the initial structure (see Figure 1c). This operation, called the
synthetic rule by Walliser and Zwirn (forthcoming) exactly corresponds to the Bayesian
conditioning rule8 under the one probabilistic level structure (see Figure 1d)9. It operates a
change of reference class but is not a true revision process (de Finetti 1949, 1972). It is
appropriate in the focusing type situation of revision (Dubois & Prade 1992; Dubois et al.
8 The two traditional equivalent specifications of Bayes’ rule are:
Form 1, “Bayes’ identity”:
€
P(carNo.1 |'goatNo.3...n' ) =P(carNo.1) × P('goatNo.3...n'|car No.1)
P('goatNo.3...n' ). This
specification can naturally be deduced from the bi-probabilistic structure of figure 1.
Form 2, “conditional probability”:
€
P(carNo.1 |'goatNo.3...n' ) =P(car No.1∧'goatNo.3...n' )
P('goatNo.3...n'). This
specification can naturally be deduced from the one-probabilistic structure of figure 1b.
€
with P('goat No.3...n') = P(car No.i)P('goat No.3...n'|car No.i)i=1
i=n
∑ = P(car No.1)P('goat No.3...n'|car No.1) + P(car No.2)P('goat No.3...n'|car No.2)
9
1996; Dubois & Prade 1997; Walliser & Zwirn forthcoming). In such focusing situation the
message concerns an object drawn at random from a population of objects (fixed universe).
For instance a clinician who knows about the distribution of the diseases and the links
between these diseases and the symptoms in a class of population and who collects the
symptoms on a patient, changes his (or her) reference class in order to focus on the cases that
share the same symptoms as the ones he (or she) has collected.
Consequently we thereby establish our second property of experimenters’
representation of MHPnn-2.
Property 2. Experimenters’ representation of revision in MHPnn-2 is a focusing situation
for which the Bayes’ rule of conditioning applies.
MHPnn-2 can be illustrated in a focusing framework as a problem of balls, urns and
meta-urns in which the message concerns an information on a ball that has been extracted
(Walliser & Zwirn forthcoming). Such problem is advantageously worded in a populational
format making explicit the bi-probabilistic level hierarchy and the situation of revision (see
below the section 4.2.4).
The urns and balls problem10: There is a meta urn No.1 that contains A urns. There are: a1
urns labeled carNo.1, ..., ai urns labeled carNo.i, ... and an urns labeled carNo.n (with
€
A = ass=1
s=n∑ ). Each a1 urns labeled carNo.1 contains n-1 balls labeled from 2 to n. Among the
A- a1 other urns, each urn labeled carNo.i contains n-1 balls labeled i. Each a2 urn labeled
carNo.2 contains n-1 balls labeled 2, ... each ai urn labeled carNo.i contains n-1 balls labeled i, ...
and each an urn labeled carNo.n contains n-1 balls labeled n. One urn is randomly extracted
from the meta urn and you must assess the probability that it is labeled urn carNo.1. Next one
ball is extracted at random from this urn and you learn the message (at layer 0) “this ball is
9 For a demonstration see Appendix 2 in Walliser & Zwirn (forthcoming)
10
not labeled 3, ... to n” (corresponding to ‘goatNo.3...n’). What is now your belief that the
selected urn is carNo.1 (P(carNo.1‘goatNo.3...n’))?
In the urns and balls problem, the message “a urn is selected and a ball is extracted, which is
not labeled 3, ... to n” is without ambiguity a focusing message. It implies that the selected
urn is not any of the urns carNo.3 , ... to carNo.n and that the probability that it is labeled carNo.1
is a1/a2(n-1) times the probability that it is labeled carNo.2.
2.4 Experimenters’ Bayesian solution (“EBS”) for MHPnn-2
Bayesian conditioning amounts to a re-standardization of the probabilities of the
remaining worlds compatible with the focusing message ‘goatNo.3...n’ (see Figure 1e). We
have:
€
P(carNo.1 |'goat No.3...n') =P(carNo.1∧'goat No.3...n')
P(carNo.1∧'goat No.3...n') +P(carNo.2∧'goat No.3...n')=
a1a1 + (n−1)a2
(2)
€
P(carNo.2 |'goat No.3...n') =P(carNo.2∧'goat No.3...n')
P(carNo.1∧'goat No.3...n') +P(carNo.2∧'goat No.3...n')=
(n−1)a2a1 + (n−1)a2
(3)
and hence the relations:
€
P(carNo.1 |"goat No.3...n")P(carNo.2 |"goat No.3...n")
= a1(n -1)a2
(4)
We get theorem 1:
Theorem 1: (i) For a2 strictly superior to the mean of set of values ai (with i > 2),
€
P(car No.1 |'goat No.3...n') < P(car No.1) = a1
A and
(ii) in the specific case where a2 corresponds to the mean of set of values ai (with
i > 2),
€
P(car No.1 | ('goat No.3...n') = P(car No.1)
10 This problem may be worded as the generalization of the traditional problem of “the three smallboxes” (Bertrand 1889).
11
Proof. By elementary calculation:
€
P(car No.1 |'goat No.3...n') ≤ P(car No.1) ⇔ a1 + (n −1)a2 ≥ aii=1
i= n
∑ ⇔ a2 ≥
aii= 3
i= n
∑(n− 2)
In particular, for standard MHPnn-2, we have11:
€
P(carNo.1 |'goat No.3...n') =1n
(5)
€
and P(car No.2 |'goat No.3...n') =1- 1n
(6)
3 Participants’ point of view
3.1 The situation of revision brings about a new hypothesis to account for MR
Although MR is commonly analyzed as evidence for the “inconsistency” of
participants, some authors have yet sketched two possible explanations for MR as a consistent
response. First, from a Bayesian point view, rules 1 and 3 are questionable because
participants’ MR could be consistent if based on subjective prior probabilities and likelihoods
different from experimenters’ beliefs12. For example, if participants believe that when the car
is behind door No.1, the host always opens doors No.3 ... to No.n (P(’goatNo.3...n’| carNo.1) =
1), then MR would correspond to a Bayesian solution13 albeit different from experimenters’.
11 In standard MHP, we have: P(carNo.1|’goatNo.3’) = P(carNo.1) = 1/3 and P(carNo.2|’goatNo.3’) = 2/3.12 For a general discussion of this point in the experimental literature on probability judgment seeBaratgin & Politzer (2006), for a general Bayesian solution to standard MHP taking into account thepossible host’s strategies see for example Wechsler, Esteve, Simonis, and Peixoto (2005). For adiscussion on the hypothesis that despite answering “MR” to standard MHP participants may actuallyhave the same representation of the task as experimenters, but assume a different host’s strategy thanthose given by implicit and explicit “rules” see for example Nickerson (1996).13 In standard MHP and standard MHPnn-2, for example the rule 3.2.1 seems intuitive because if thehost has originally no preference for a specific door when he/she initially places the car, he/she isexpected to have no preference when choosing to open n-2 doors with goats behind. Yet, it seemsmore debatable in MHP and MHPnn-2 versions because initially the host has a preference on the doors(unequal prior probabilities).
12
Second, one can question Postulate 1 that considers message ‘goatNo.3...n’ as certain14 and
Property 1 that forces to stick to the additivity framework15. However these two explanations
actually appear limited, as participants seem to adhere to experimenters’ Property 1 in the
numerous experiments on standard MHP and MHP versions. When participants are indeed
explicitly required to state their prior probabilities, they actually share experimenters’ priors
(Ichikawa & Takeichi 1990; Baratgin & Politzer 2003; Franco-Watkins et al. 2003). Besides
when experimenters explicitly specify the strategy of the host, participants nevertheless give
MR (Shimojo & Ichikawa 1989; Ichikawa & Takeichi 1990; Granberg & Brown 1995;
Krauss & Wang 2003; Burns & Wieth 2004)16. Consequently an explanation for MR still
needs to be found.
Let’s next analyze whether participants adhere to experimenters’ Property 2. The
relevant literature shows that experimenters who regard MR as erroneous generally fail to
consider that participants’ interpretation of the host’s message might not conform to the only
situation they envisage, namely a focusing situation of revision. Yet participants in MHPnn-2
can infer pieces of information that differ from those meant by the experimenter. This
approach is new to psychological research because, other situations of revision than focusing
have scarcely ever been studied (for review, see Baratgin & Politzer 2007a). In few studies,
participants learn a message that specifies or invalidates an initial belief regarding a universe
also considered as fixed. In this situation, called “revising”, Bayes' rule remains the adequate
rule (among other possible ones) for an individual to change his (or her) degrees of belief.
14 In the situation where the message would be dubious, Jeffrey’s rule would be adequate (for asolution to standard MHP in this situation see Loschi et al. 2006).15 This postulate is traditional in the field of studies on probability judgment. It assumes thatparticipants’ degrees of belief are quantitative and obey to Kolmogorov’ s axioms. However theplausibility of this hypothesis is debatable (Baratgin & Politzer 2006). Some authors have proposed asolution for standard MHP in a formalism different from traditional additive probability (see Appendix2).16 However no experimental study analyzes whether participants have a precise distribution ofprobability between layers 1 and 0. In this case, participants’ representation of MHP would be a“distribution of events” (see Appendix 2 on this question).
13
For instance in the urns and balls problem the message “there are no balls labeled 3, ... to n in
the extracted urn” is a revising message that changes our belief into the certitude that it is urn
carNo.2 that has been extracted. This operation amounts to eliminating the possible urns with
balls labeled 3, ... to n and changes our beliefs on the remaining urns (here the only urn
carNo.2) according to Bayes’ rule. Such process appropriate to a bi-probabilistic structure is
called “maximal rule” (Walliser & Zwirn forthcoming). In contrast to the focusing and
revising situations, individuals may also have to revise their degrees of belief when the initial
universe - now considered as dynamic - has changed according to the information conveyed
by the new message. Such situation, called “updating”, has a clear definition supported by
both a set of axioms for belief revision (Katsuno & Mendelzon 1992) and for probability
revision (Walliser & Zwirn 2002). In the updating situation of revision, the message specifies
a change of the universe, which is considered as time evolving. The updating message is easy
to illustrate when it stems from an action or an intervention that modifies the initial
knowledge of the original universe. Some possibilities may be removed by an exterior action.
But it is noteworthy that updating may occur without an explicit intervention. This would be
the case with the outbreak of a new disease that transforms the knowledge base of physicians,
or with new statistical data that radically alter a demographic model. In the urns and balls
problem, an updating message at layer 0 would correspond to the message “the balls that are
labeled 3, ... to n have been removed from all urns” (in MHPnn-2 this message is equivalent to
“the doors numbered 3, .. to n (with goats behind) has been removed” and noted carNo.3...n).
The updating message indicates that the original problem has evolved.
You are now faced with a meta urn No.1 that contains that contains A urns. There are:
a1 urns labeled carNo.1, ..., ai urns labeled carNo.i, ... and an urns labeled carNo.n (with i = 1, ... to
n and
€
A = ass=1
s=n∑ ). The a1 urns labeled carNo.1 contain only one ball labeled 2, the a2 urns
carNo.2 always contain n-1 balls labeled 2 and each other urns carNo.i (i = 3, ... to n) are now
14
empty. It implies that the selected urn is not any of the urns carNo.3 , ... to carNo.n and that the
probability that it is labeled carNo.1 is a1/a2 times the probability that it is labeled carNo.2. This
result corresponds to MR in MHPnn-2. The operation amounts to removing balls labeled 3 to n
from the possible urns and to changing their composition accordingly by Bayes’ rule. The
updating rule appropriate in a bi-probabilistic structure is called “minimal rule (Walliser &
Zwirn forthcoming) (see section 3.3.2). A correspondence exists between the situations of
revision at different layers (see for a generalized analysis Walliser & Zwirn, forthcoming). In
the urns and balls problem, after the updating message at layer 0 “the balls that are labeled 3,
... to n have been removed from all urns” , induces the focusing message at layer 1 (urns
level): “the extracted urn from the large urn is not a urn of the type carNo.3, ... to carNo.n” (that
is equivalent in MHPnn-2 to message “the doors No.3, ... to No.n hide goats” noted
goatNo.3...n)17. Psychologists have never explicitly considered the situation of updating in their
numerous investigations. Studying this situation however appears necessary for a better
understanding of human probability revision (see for a review Baratgin & Politzer 2006).
Even though experimenters assume or instruct participants that the universe is fixed, it may be
the case that participants yet have a dynamic interpretation of it. So, when assessing the
coherence of participants’ judgment, experimenters must be aware of this possibility and
ready to use the appropriate formalism. A wrongful participants’ response given in a focusing
situation (for which Bayes’ rule is the only adequate measure) may actually reveal correct if
participants interpret the situation in an updating framework (for which Bayes’ rule does not
apply) and the response is consistent with the adequate revision rule for updating. We argue
that MR may be seen as consistent if participants typically infer the focusing message
‘goatNo.3...n’ as the updating message carNo.3...n.
17 This situation is isomorphic to a problem with two sorts of urns; the urns carNo.1 that contain only one balllabeled 2 and the urns carNo.2 that contain n-1 balls labeled 2 Thus the updating message is also equivalent to anupdating message at layer 1 (urns level): “the urns No.3, ..., No.n have been removed from the meta urn”.
15
3.2 A pragmatic explanation for the updating representation
The hypothesis that participants interpret MHPnn-2 as an updating situation of revision
can be supported by pragmatic analysis18.
In the two first stages (i) and (ii) of MHPnn-2, participants are required to pick a door
that hides a car among n doors presented in front of them. They make their initial choice of a
door (here No.1) without any special information related to this object. Because of
uncertainty participants attach a prior probability to each door (here explicitly given by
experimenters). In the framework of relevance theory (Sperber & Wilson 1995), the
communicative principle posits that any utterance conveys a presumption of its own (optimal)
relevance. In other words, participants assume that if the experimenter sends a message, it
deserves to be treated. Thus for participants any new information given by the experimenter
carries a presumption of importance in relation with the aim of the problem (to know where
the car is). In fact, participants can expect the new information to be highly important
(relevant) as they have initially chosen the door randomly and this is mutual knowledge
shared with the experimenter. In addition according to relevance theory, participants are more
inclined to pay attention to messages with important contextual effect and low cognitive effort
of treatment. The focusing message ‘carNo.3...n’ corresponds to stages (iii) and (iv). At stage
(iii), the host warns that he will open n-2 doors (out of the n-1 remaining doors) that hide
goats. This message actually comprises two compacted pieces of information. The first
element is implicit. It corresponds to the host’s strategies to open n-2 doors among doors
No.2, ... to No.n that hide goats. It constitutes the key to construct the representation of the
puzzle (Figure 1a). The second element refers to the host’s explicit declaration to open n-2
doors among doors No.2, ... to No.n that hide goats. The comprehension of the first element
18 Some studies on probability judgment have supported the hypothesis along which participants infera different representation of the task than experimenters do for various pragmatic reasons (see for
16
requires a sizeable cognitive effort because participants must work out the n-2 complementary
possibilities that the host may show goats behind doors (No.2, No.4, ... and No.n), ..., or
(No.2, ... and No.n-1) then keep this in memory. The second element of the message requires
low a cognitive effort of treatment and is contextually powerful and expected (so would be
the case of the n-2 other possible messages). It carries a strong presumption of importance in
relation with the aim of the problem (to know where the car is). Participants induce that “n-2
of the n-1 doors (No.2, ..., No.n) will be eliminated”. This should lead them to favor the two
remaining doors (n-2 doors are removed as well as the possibilities to have a car behind).
Participants nevertheless take into account the host’s implicit strategies and expect following
the host’s declaration to be promptly confronted with a “new simplified” problem with two
doors (No.1 and an other door) to which they may attach a new distribution of probability.
The hypothesis that participants are in expectation of an evolving problem was tested in
Baratgin (1999; 2004) with a revised version of standard MHP limited to three stages. After
the first two stages (i) and (ii), the host delivers at stage (iii) an uninformative message (that is
the already known message of stage (i)): “at least one door (out of No.2 and No.3) has a goat
behind”. A redundancy effect was evidenced as a sizeable number of participants nevertheless
modified their initial degrees of belief and give in a large majority MR. In other words
although a message may actually be uninformative, it still may prompt participants to infer
information from it and to revise their perception of the problem into a simplified problem
with two doors (No.1 and an other door). Now, when participants receive the message
(resulting from the host’s action) “the car is not behind doors No.3, ...to No.n”. They
consider it as the answer to the expectation to see n-2 possible doors removed. The message
benefits from an overwhelming contextual effect and amounts in participants’ view to
concretely eliminating doors No.3, ..., No.n from the set of possible doors (carNo.3...n).
instance Dulany & Hilton 1988; Krosnick et al. 1990; Politzer & Noveck 1991; Macchi 1995; Politzer
17
Participants are then left with two doors (door No.1 and door No.2) they nevertheless adhere
to Property 1 (the bi-probabilistic level structure). Consequently, participants are strongly
invited to increase their beliefs related to doors No.1 and No.219. From an experimental point
of view the analysis of the verbal responses of the majority of participants in standard MHP is
consistent with this pragmatic explanation: “we have shifted from a 3-doors problem to a 2-
doors problem after the host’s action ” (Baratgin & Politzer 2003).
Overall we conclude that participants interpret standard MHP as an updating problem
of revision and that “EBS” violates expectations resulting from ostensive communication.
The focusing situation does not really correspond to a dynamic process of probability
revision because in a focusing situation there are no truly invalidated worlds. On receiving the
message ‘goatNo.3...n’, participants should change the reference of the whole initial set of
possible worlds (see section 2.3) but stick to the same probability distribution P. This is
explicit in the urns and balls problem but on the contrary, the statement of MHPnn-2 suggests
an updating situation of revision. At time t0 participants are invited to work out their prior
distribution on n doors (layer 1); at time t1 they integrate the new message resulting from the
host’s action (layer 0). This action modifies perceptibly the problem (two doors). Then the
experimenter asks participants a judgment of probability that can be interpreted as a true
revision of their initial probability distribution P leading to a updating revision. This hint
contained in the statement deters participants from envisaging the focusing situation.
3.3 MHPnn-2 in an updating interpretation
3.3.1 The general Imaging rule
& Macchi 2005).19 Hence “EBS” contradicts the principle of relevance in numerous cases (see Theorem 1). Notably in allexperiments where experimenters have proposed a version of MHP where a2 > a3 or a standard version ofMHPnn_2. “EBS” neither increases (nor changes for standard MHPnn-2) participants’ initial degrees of beliefattributed to carNo 1. The message ‘carNo.3...n’ would only lead participants to increase their initial degrees of
18
In MHPnn-2 if the message ‘goatNo.3...n’ is interpreted (albeit mistakenly) as the
updating message “the doors No.3, ...to No.n have been removed by the host” ; it translates a
physical operation on the system, modifying its initial structure. In this situation, the
adequate method of revision for a one probabilistic level is different and Lewis’ rule (or
Imaging) (Lewis 1976; Gärdenfors 1988; Lepage 1997) reveals the appropriate operation of
revision (see the Imaging generalized rule Walliser & Zwirn 2002). . Generally Bayesian
conditioning rule and imaging yield different solutions20. Imaging causes a revision of the
prior probabilities on the possible worlds in such a way that the posterior probabilities are
obtained by shifting the initial probabilities of invalidated worlds to newly validated worlds.
Each invalidated world sees its probability attributed to the closest set of worlds that remain.
One difficulty of Imaging lies in defining the notion of “closest” world. Hence the few
authors who have envisaged an imaging solution to standard MHP have proposed two
different possible distances (Dubois & Prade 1992; Cross 2000). Yet these authors have
considered an Imaging operation with regards to the compacted probabilistic level of Figure
1b. Under this compacted representation there is no clue to choose a particular distance. Let
€
P'goat No.3...n'
+ (carNo.1) and
€
P'goat No.3...n'
+ (carNo.2) be the probabilities obtained by Imaging after the
belief attached to carNo.2. These results can be interpreted as a violation of logic (see Dubois & Prade 1992, for adiscussion in standard MHP).20 It is easy to see that the difference with Bayes’ rule lies in the redistribution of the invalidated world(carNo.1∧’goatNo.2’). Indeed Bayes’ rule can also be seen as a redistribution of invalidated worldsproportionally to prior probability of remaining worlds. Consequently “BS” can be formulated asfollows:
€
P(carNo.1 |'goatNo.3...n' ) =P(carNo.1∧'goatNo.3...n')
P(carNo.1∧'goatNo.3...n') + P(carNo.2∧'goatNo.3...n' )
€
= P(car No.1∧'goat No.3...n')
+P(carNo.1∧'goat No.3...n')
P(car No.1∧'goat No.3...n') + P(car No.2∧'goat No.3...n' )1− P(car No.1∧'goat No.3...n') - P(car No.2∧'goat No.3...n' )( )
€
= P(carNo.1∧'goatNo.3...n')
+P(carNo.1∧'goatNo.3...n')
P(carNo.1∧'goatNo.3...n') + P(carNo.2∧'goatNo.3...n' )P(carNo.1∧'goatNo.2.4...n') + ...+ P(carNo.3∧'goatNo.2.4...n' )( )
For a detailed axiomatic analysis of this difference see Walliser & Zwirn (2002)
19
updating message ‘goatNo.3...n’.
€
P'goat No.3...n'
+ (carNo.1) is called the image of P(carNo.1) on worlds
where ‘goatNo.3...n’ is true. When the updating message ‘goatNo.3...n’ is learnt, the probabilities
concerning the 2(n-2) invalidated worlds {carNo.1∧’goatNo.2.4…n’, carNo.1∧’goatNo.2.3.5.6…n’,
..., carNo.1∧‘goatNo.2…n-1’, carNo.3∧’goatNo.2.4…n’, ..., carNo.n∧‘goatNo.2…n-1’} are
redistributed over the other possible worlds that participants consider as the closest. The n-2
worlds {carNo.1∧’goatNo.2.4…n’, carNo.1∧’goatNo.2.3.5.6…n’, ..., ..., carNo.1∧‘goatNo.2…n-1’}
are incompatible with the world carNo.2 but perfectly compatible with the world carNo.1. Thus
it is natural to assume that the closest world to these n-2 worlds is carNo.1 and that the weight
P(carNo.1∧’goatNo.2.4…n’) + P(carNo.1∧’goatNo.2.3.5.6…n’) +...+ P(carNo.1∧‘goatNo.2…n-1’) is
reallocated over the single world carNo.1 (Dubois & Prade 1992; Cross 2000). As far as the
invalided worlds {carNo.3∧’goatNo.2.4…n’, ..., carNo.n∧‘goatNo.2…n-1’} corresponding to n-2
worlds carNo.i (with i = 3, ..., i = n, cf. relation 1) are concerned, there are different
possibilities to define their “closest” world. In a general Imaging (Gärdenfors 1988) the
weight of the invalidated worlds are redistributed over several remaining worlds. In MHPnn-2
each weight of the invalidated worlds, P(carNo.i) (with i = 3, ..., i = n) can be redistributed on
the two remaining worlds carNo.1 and carNo.2.
Let βi be the proportion of weight P(carNo.i) reallocated to world carNo.1 (for i = 3, ..., i
= n, βi ∈ [0, 1]). The general Imaging rule on considering the one probabilistic level gives:
€
P'goat No.3...n'
+ (car No.1) = P(car No.1∧'goat No.3...n') +P(car No.1∧'goat No.2.4...n') +
... + P(car No.1∧'goat No.2...n -1') + βiP(car No.i)i=3
i=n
∑
€
P'goat No.3...n'
+ (carNo.1) = P(carNo.1) + β iP(carNo.i)i=3
i=n
∑ (7)
and similarly
€
P'goat No.3...n'
+ (carNo.2) = P(carNo.2) + (1−β i )P(carNo.i)i=3
i=n
∑ (8)
We thus establish theorem 2 for general imaging:
20
Theorem 2. (i) If there is at least one i with βi ≠ 0 then
€
P'goat No.3...n'
+ (carNo.1) > P(carNo.1)
and (ii)
€
if β i = 0, for i = 3, ..., i = n, then P'goat No.3...n'
+ (car No.1) = P(car No.1) .
However, as we have seen in section 2.2, the belief representation of MHPnn-2
corresponds to the bi-probabilistic level of Figure 1a and the one-probabilistic level of Figure
1b is only a consequence of this bi-probabilistic level. On considering the bi-probabilistic
level it is now possible to detail the appropriate redistributions of the weight of invalidated
worlds carNo.i (with i = 3, ..., i = n) and thus define (βi), i = 3, ..., i = n.
3.3.2 The “proportional Imaging rule” gives MR in MHPnn-2
In a general analysis of the revision rules for bi-probabilistic level, Walliser & Zwirn
(1997; forthcoming) propose the “minimal rule” as the appropriate updating rule. It consists
in revising in a Bayesian way the distributions of level 1 that are compatible with the
message, the other distributions being abandoned with a necessary homothetic adjustment of
the distribution at level 2 (see Figure 2a).
-------------------------------------------------------------
Insert Figure 1 about here
-------------------------------------------------------------
This rule is obvious in the urns and balls problem (see section 3.1). We call this rule
the “proportional Imaging rule ” because it corresponds (in the probabilistic structure of one
level) to the reallocation of the weight of incompatible worlds over the remaining worlds
proportionally to their prior probabilities (see Figure 2b). In MHPnn-2, the weight P(carNo.i)
(for i = 3, ..., i = n ) must be reallocated over the worlds carNo.1 and carNo.2 proportionally to
their prior probabilities. In this case
€
βi =P(carNo.1)
P(carNo.1) +P(carNo.2) and is independent from i.
21
The solution of MHPnn-2 by this proportional Imaging rule (from now on “PIS”) leads to the
same result as the revision by Bayes’ rule after the simpler focusing message goatNo.3...n (see
also the urns and balls problem in section 3.1). Hence we have the following theorem 3.
Theorem 3. The proportional Imaging rule is mathematically isomorphic to a Bayesian
conditioning on the “naive set” {carNo.1, ..., carNo.n}21.
Proof.
€
P'goat No.3...n'
+ (car No.1) = P(car No.1) + P(car No.1)P(car No.1) +P(car No.2)
× P(car No.i)i= 3
i= n
∑
= P(car No.1) 1+P(car No.i)
i= 3
i= n
∑P(car No.1) +P(car No.2)
€
P'goat No.3...n'
+ (carNo.1) =P(carNo.1)
P(carNo.1) +P(carNo.2)= P(carNo.1|goat No.3...n) =
a1a1 +a2
(9)
Similarly:
€
P'goat No.3...n'
+ (carNo.2) =P(carNo.2)
P(carNo.1) +P(carNo.2)= P(carNo.2|goat No.3...n) =
a2a1 +a2
(10)
and we have the relation:
€
P'goat No.3...n'
+ (carNo.1)
P'goat No.3...n'
+ (carNo.2)= P(carNo.1)P(carNo.2)
=a1a2
(11)
In particular, for standard MHPnn-2, we have:
€
P'goat No.3...n'
+ (carNo.1) = P'goat No.3...n'
+ (carNo.2) =12
(12)
It is therefore obvious that:
€
P'goat No.3...n'
+ (car No.1) > P(car No.1) and P'goat No.3...n'
+ (car No.2) > P(car No.2) .
Solutions (9) and (10) exactly correspond to MR in MHP22 (Ichikawa 1989; Ichikawa
& Takeichi 1990; Granberg 1996; Yamagishi 2003). Proportional Imaging verifies theorem 2
21 We make a parallel here with the remark of some theorists who have interpreted MR as the result ofa Bayesian conditioning on a “naive set” while “EBS” is the result of a Bayesian conditioning on amore “sophisticated set” (Jeffrey 1988; Grünwald & Halpern 2003). In Appendix 3, we show that theproportional Imaging rule is equivalent to Pearl (2000)’ s Bayesian conditioning after one intervention(designed by Pearl as the “do” operator).22 In standard MHP, we have:
€
P'goatNo.3' ( carNo.1) = P'goatNo.3' ( carNo.2) =12
.
22
(i) and supports the pragmatic argument of section 3.2. It makes a good candidate to
experimentally describe and theoretically gauge participants’ judgment (see also section 4).
Finally “PIS” verifies the below theorem 4.
Theorem 4. For any n > 2,
€
P'goat No.3...n'
+ (carNo.1) > P(carNo.1 |'goat No.3...n') and
€
P'goat No.3...n'
+ (car No.2) < P(car No.2 |'goat No.3...n') .
Proof. It is obvious because for n > 2;
€
a1
a1 +a2
> a1
a1 +(n -1)a2
and
€
(n -1)a2a1 +(n -1)a2
>a2
a1 +a2
3.3.3 The “egalitarian Imaging rule”
It is noteworthy that when a1 is equal to a2, the proportional Imaging rule is isomorphic to the
egalitarian Imaging rule23, which redistributes on the worlds carNo.1 and carNo.2 half of the
weight of the cancelled worlds carNo.i (with i = 3, ..., i = n). The equidistance of the two
worlds carNo.1 and carNo.2 to worlds carNo.i has been suggested by Dubois and Prade (1992) in
standard MHP, and may be seen as the result of the n possible worlds being initially equally
likely. Let
€
P''goat No.3...n'
+ (carNo.1) be the revision of P(carNo.1) with this “egalitarian Imaging rule
” (we use P’ to differentiate from proportional Imaging rule). The general solution by
egalitarian Imaging rule in MHP nn-2 (from now on “EIS”) gives:
€
P''goat No.3...n'
+ (carNo.1) =a1A
+12×A- a1 − a2
A=A +a1 − a22A
=121+ P(carNo.1) - P(carNo.2)( ) (13)
and
€
P''goat No.3...n'
+ (carNo.2) =A- a1 + a22A
=121+ P(carNo.2) - P(carNo.1)( ) (14)
Thus we have the relation:
€
P''goat No.3...n'
+ (carNo.1)
P''goat No.3...n'
+ (carNo.2)= 1+P(carNo.1) - P(carNo.2)1+P(carNo.2) - P(carNo.1)
=1+a1 − a21+a2 − a1
(19)
23 The term Egalitarian Imaging rule is borrowed from Walliser & Zwirn (2002). This rule is alsoanalyzed in Lepage (1997).
23
In the case of standard MHPnn-2 the solution gives MR:
€
P''goat No.3...n'
+ (carNo.1) = P' 'goat No.3...n'+ (carNo.2) =
1n
+12×(n - 2)n
=12
(20)
3.3.4 The “simple Imaging rule ”
As suggested by Cross (2000) in standard MHP a second distance can be assumed in
the probabilistic structure of one level. Such distance considers that the worlds carNo.i (with i
> 1) are closer to one another rather than to carNo.1. In this case, the weights P(carNo.i) are
reallocated on the sole world carNo.2. Intuitively this distance can be understood as if the
weight of the worlds carNo.2∨...∨carNo.n should not change after the updating message
‘goatNo.3 ... .n’. This distance seems artificial with the bi-probabilistic level of Figure 1a but
can be analyzed as the appropriate rule in a different representation (with a supplementary
level) where the different doors No.i (with i >1) are gathered under a similar probability. For
instance the door No.1 is white and all others doors are black (see Figure 2c). In a focusing
framework this tri-probability level structure is isomorphic to the original bi-probabilistic
level but in an updating framework the minimal rule corresponds to the simple Imaging rule.
Let
€
P"'goat No.3...n'
+ (carNo.1) be the revision of P(carNo.1) with this “simple Imaging rule ”
(we use P’ to differentiate from proportional Imaging rule).
€
P"'goat No.3...n'
+ (car No.1) = P(car No.1) = a1
A (21)
€
P"'goat No.3...n'
+ (car No.2) = P(car No.2) + P(car No.i)i=3
i=n
∑ = P(car No.i) =i=2
i=n
∑ A− a1
A (22)
Thus we have the relation:
€
P"'goat No.3...n'
+ (carNo.1)
P"'goat No.3...n'
+ (carNo.2)= P(carNo.1)
P(carNo.i)i=2
i=n
∑=
a1A - a1
(23)
The solution (from now on “SIS”) applied to MHPnn-2 verifies theorem 2 (ii).
More generally, we have the following theorem 4 (derived from theorem 1):
24
Theorem 5. ‘SIS” corresponds to “EBS” in MHPnn-2 in the specific case where a2
corresponds to the mean of the set of values ai (with i = 3, ..., = n)
In standard MHPnn-2, “SIS” corresponds to “EBS”:
€
P"'goat No.3...n'
+ (carNo.1) = P(carNo.1 |'goat No.3...n') =1n
and (24)
€
P"'goat No.3...n'
+ (carNo.2) = P(carNo.2 |'goat No.3...n') =1n
+n − 2n
=n −1n
(25)
4 Experimental hindsight
4.1 The updating interpretation underlies cognitive explanations for MR
Several cognitive explanations to account for MR have been put forward in
psychological experimental studies conducted on MHP and standard MHP as well as standard
MHPnn-2. They do not reveal independent one from the others and appear to be actually under
lied by the updating interpretation of ‘goatNo.3...n’. In this section we will review three main
families of such cognitive arguments. We will adopt the notation DB as apposed to P to
represent participants’ degree of belief in order to stress that we focus here on cognitive
processes rather than on theoretical relationships.
4.1.1 The heuristics explanations
The first set of explanations proposes that participants use “simple heuristics” or
“subjective heuristics” to form their judgments of probabilities (Ichikawa 1989; Shimojo &
Ichikawa 1989; Ichikawa & Takeichi 1990; Falk 1992; Yamagishi 2003).
4.1.1.1 The heuristics explanations for MR
According to these authors, participants who give MR in standard MHP use the two
following heuristics:
25
“The number of cases heuristic” (NC): When the number of possible alternatives is N, the
probability of each alternative is 1/N.
“The constant ratio heuristic” (CR): When one alternative is eliminated, the ratio of
probabilities for the remaining alternatives is equal to the ratio of respective prior
probabilities.
If participants apply NC or CR to standard MHPnn-2, they will cut down the number of
alternatives from n doors to two when learning the message ‘goatNo.3...n’. They will thus
conclude that: DB(carNo.2|’goatNo.3...n’) = DB(carNo.1|’goatNo.3...n’) = 1/2. It proves difficult to
know what strategy participants actually implement to answer 1/2 in standard MHPnn-2 since
NC and CR are undistinguishable. However, when Shimojo and Ichikawa (1989) and
Ichikawa and Takeichi (1990) propose MHP to their participants a large majority of
participants actually use CR. This leads us to theorem 6.
Theorem 6. “EIS” given by the egalitarian Imaging rule is isomorphic to the solution for
standard MHPnn-2 given by NC. More generally “PIS” given by the proportional Imaging rule
is isomorphic to the solution for MHPnn-2 based on CR.
Proof. We know by the Rule 1 that DB(carNo.i) = P(carNo.i) for i = 1, ..., i = n. CR definition
corresponds to relation (11):
€
DB(carNo.1 |'goat No.3...n')DB(carNo.2 |'goat No.3...n')
=P(carNo.1)P(carNo.2)
=a1a2
, and by the
complementarity constraint of additive probability (Property 1) we have:
€
DB(carNo.1 |'goat No.3...n') +DB(carNo.2 |'goat No.3...n') =1, thus:
€
P(car No.1)P(car No. 2)
+1
DB(car No.2 |'goat No.3...n') =1 and we find the relations (9) and (10):
26
€
DB(car No.1 |'goat No.3...n') = P(car No.1)P(car No.1) +P(car No.2)
= P'goat No.3...n'
+ (car No.1) and
€
DB(carNo.2 |'goat No.3...n') =P(carNo.2)
P(carNo.1) +P(carNo.2)= P
'goat No.3...n'
+ (carNo.2)
4.1.1.2 The heuristics explanations for “EBS” in standard MHP
In parallel Shimojo and Ichikawa (1989, p. 6-7) propose a third heuristic for the thin
minority of participants in standard MHP, who give “EBS” (see also Falk, 1992). These
authors show that these participants do not actually use Bayes’ rule to give “EBS” but use
rather the following heuristic:
“The Irrelevant, Therefore Invariant heuristic” (ITI): “If it is certain that at least one of
the several alternatives (A1, A2, …, Ak) will be eliminated. When the information specifying
what alternatives to be eliminated is given, it does not change the probability of the other
alternatives (Ak+1, Ak+2, …, An)“.
Participants who use this heuristic do not modify their prior probability P(carNo.1) after
the message ‘goatNo.3...n’ just as “SIS”.
Theorem 7. “SIS” given by the simple Imaging rule is isomorphic to the solution for MHPnn-2
given by ITI.
Proof. Contrary to CR and NC, the definition of ITI does not predict how the probabilities of
the remaining alternatives must change once some have been removed. Nevertheless, using
the basic complementarity constraint of the axiom of additivity of probabilities (Property 1),
we can suppose that:
€
DB(carNo.2 |'goat No.3...n') =1-DB(carNo.1 |'goat No.3...n') =1- P(carNo.1) = P(carNo.i)i=2
i=n
∑ and we
find again relations (21) and (22) of “SIS”.
27
4.1.1.3 The “Subjective Superior heuristic” (SS):
Finally Shimojo and Ichikawa (1989) assume that the answers based on the three
above mentioned heuristics actually result from a “Subjective Superior heuristic” defined as
participants’ following belief.
The “Subjective Superior heuristic”: “The posterior probabilities of the remaining
alternatives should never be less than their priors when one alternative is removed”.
This belief is incompatible with theorem 1 and more generally with the process of
conditioning following Bayes’ rule (see for a demonstration Ichikawa 1989). However it is
compatible with theorem 2 corresponding to an updating situation.
4.1.2 The Mental Models Theory
The second main explanation for MR was proposed by Johnson-Laird et al. (1999)
based on their Mental Models Theory (from now on noted MMT) (see also Girotto &
Gonzalez 2000). According to these authors, participants are not able to construct the
complete set of mental models the puzzle entails because they have a limited working
memory. MMT in probability situation can be described by four principles.
(i) The “truth principle”: Participants first represent the probability problem by
constructing the sets of “mental models” (representations of the different possibilities), which
probabilities are strictly superior to 0, (ii) the “equiprobability principle”: Participants
construct a diagram of equiprobable alternatives (iii) The “proportionality principle”:
Participants assess the probability of an event E proportionally to the number of mental
models (nE) in which event E occurs, that is to say nE/n with n the number of mental models
and (iv) the subset principle: Granted equiprobability, a conditional probability, P(E| B),
depends on the subset of B that is E, and applying the proportionality principle to E and B
yields the numerical value.
28
We notice that the principles (i), (ii) and (iii) actually make up an algorithm to
construct the experimenter’s representation of Figure 1b in a populational format (see section
4.2.4). Meanwhile the subset principle (iv) corresponds to a natural translation of a focusing
procedure24. Hence MMT can be interpreted as an alternative normative method to describe
MHPnn-2 (mental models corresponding to possible worlds).
For instance, the description of standard MHP in terms of mental models involves 6
models: 2 models when the car is behind the chosen door (door No.1) because the host then
either exhibits a goat behind door No.2 (first model) or behind door No.3 (second model).
Hence we have the following models
€
car No.1 'goat No.2'
car No.1 'goat No.3'
. The equiprobability principle,
assumes that 2 identical models exist in the case where the car is behind door No.i (i > 1) and
the host shows that door No.j (with j > 1 and j ≠ i) hides a goat. Hence we have the additional
4 models:
€
car No.2 'goat No.3'
car No.2 'goat No.3'
car No.3 'goat No.2'
car No.3 'goat No.2'
.
Now when the host shows that door No.3 hides a goat, the “subset principle” retains
the 3 models that contain ‘goatNo.3’:
€
car No.1 'goat No.3'
car No.2 'goat No.3'
car No.2 'goat No.3'
and by the “proportionality
principle” we arrive at “EBS”.
However, Johnson-Laird et al. (1999), outline that when the car is behind door No.1,
for working memory limitation, participants simply think: “the host chooses another door”.
Similarly, participants fail to construct the identical models in each case when the car is not
24 These principles actually rediscover Laplace’s (1986, p. 38-39) traditional first two principles: (i) “Theprobability is the ratio of the number of favorable cases to the number of all possible cases”; (ii) but this holdsunder the assumption that the various cases are equally possible; if they are not…. the probability will be thesum of the possibilities of each possible case” (Author’s translation).
29
behind door No.1. They consequently built only the 3 following mental models:
€
car No.1 other door
car No.2 other door
car No.3 other door
. When the host shows that door No.3 hides a goat, the use of the
“subset principle” implies that participants consider only the 2 mental models compatible
with this message:
€
car No.1 other door
car No.2 other door
and by the proportionality principle it gives MR.
It is obvious that manipulating mental models requires significant a memorial
cognitive effort (
€
(n−1) aii=1
i= n
∑ models for MHPnn-2). TMM’s explanation would hold if
participants effectively use mental models to form their probability judgments. However this
hypothesis is not widely spread (see for example Brase 2002). In MHP nn-2, participants seem
to accept Property 1 of experimenters’ representation (see section 3.1) thus confirming a full
representation of the puzzle. Participants can also use a representation different from mental
models to understand MHP nn-2. For instance, the formal bi-probabilistic structure described in
section 2.2 (see Figure 1) requires less effort of manipulation in memory than a description in
terms of mental models. Nevertheless, if one accepts the principles of TMM, an alternative
explanation for MR can be again formulated in terms of the updating interpretation of revision
situation. If participants interpret the message ‘goatNo.3...n’ in an updating framework, the
“subset principle” (iv) no longer holds. The problem has been modified and participants must
again (as if answering to a new problem) construct a new set of mental models. In standard
MHP, the “truth principle” induces a new set of 2 mental models:
€
car No.1 goatNo.2
car No.2 goatNo.1
.
Participants by the “equiprobability principle” find MR. This process corresponds to the
proportional Imaging rule in a MHPnn-2 formulated in a populational format.
4.1.3 The proportional Imaging rule explains the partition-edit-count strategy
30
A third explanation for MR was proposed by Fox and Levay (2004) (see also Fox &
Rottenstreich 2003; Fox & Clemen 2005).
Participants evaluate conditional probabilities in (i) subjectively partitioning the sample
space into a set of n interchangeable events suggested by the statement of the problem, (ii)
editing out events that can be eliminated by the new message and (iii) counting the number
of remaining events and reporting the ratio.
However these authors think that partitioning correctly the sample space in standard
MHP proves difficult because participants only focus on the sample space most obviously
suggested by the statement. In this view, participants in MHPnn-2 do not envisage the correct
experimenters’ set (i.e. 2(n-1) possible worlds {carNo.1∧’goatNo.2.4…n’,
carNo.1∧’goatNo.2.3.4.6…n’, ..., carNo.1∧’goatNo.2…n-1’, carNo.1∧’goatNo.3…n’,
carNo.2∧’goatNo.3…n’, carNo.3∧’goatNo.2.4…n’, ..., carNo.n∧’goatNo.2…n-1’} of Figure 1b) but
the limited sample space of n possible worlds {carNo.1, ..., carNo.n}. Hence, applying (ii) and
(iii) on this naïve space gives MR. As we have seen in relations (9) and (10) this cognitive
explanation supports further our hypothesis of an updating interpretation of the message of
revision. Indeed the updating interpretation of message ‘goatNo.3...n’ forces participants to
consider only the simpler probabilistic level corresponding to this naive set (see section
3.2.1).
4.2 “EBS” is facilitated when the format of MHPnn-2 enhances the focusing interpretation
In the experimental literature some studies show that specific formats of MHPnn-2’
statement seem to prompt a higher rate of “EBS” in participants’ response. These results have
triggered some psychological debates on the underlying cognitive processes at play (Barbey
& Sloman Forthcoming). However our binary view of MHPnn-2, in terms of experimenter’s
“correct” focusing and participants’ “misleading” updating interpretation of the situation of
31
revision, invites to a new reading of these experimental results. It leads us to conclude that the
different formats that facilitate “EBS” actually enhance the focusing understanding of the
situation of revision.
4.2.1 Natural frequencies format
Some authors argue that the human mind would have developed in its evolution a
module designed to process natural frequencies acquired by natural sampling (Cosmides &
Tooby 1996; Gigerenzer & Selten 2001). Various studies prove that probability judgments
appear more Bayesian when experimental paradigms refer to natural frequencies (as opposed
to single-event probabilities) (Gigerenzer & Hoffrage 1995; Cosmides & Tooby 1996; Brase
et al. 1998; Hoffrage et al. 2002). This result is confirmed in problems written in a natural
frequencies format isomorphic (from a mathematical point of view) to standard MHP (Tubau
& Alonso 2003; Yamagishi 2003). For instance the following problem used in Yamagishi
(2003)
The Gemstone problem (natural frequencies format): A factory manufactures 1200
artificial gemstones daily. Among the 1200, 400 gemstones are blurred, 400 are cracked, and
400 contain neither. An inspection machine removes all cracked gemstones, and retains all
clear gemstones. However, the machine removes half of the blurred gemstones. How many
gemstones pass the inspection, and how many among them are blurred?
In this natural frequencies format, the bi-probabilistic structure illustrated by Figure 1
in section 2.2 is made explicit and unambiguous. To draw the parallel with standard MHP, we
respectively have blurred, cracked, and neither that correspond to carNo.1, carNo.3 and carNo.2
while ‘goatNo.3’ corresponds to the inspection process. However this format is different from a
pragmatic point of view. The three pragmatic arguments inherent to MHPnn-2’s three stages
(see section 3.2) to explain the updating interpretation, no longer exist. Participants know that
32
the situation of revision is static and that they must stick to their original distribution P
(explicit in the statement). The message not cracked equivalent to ‘goatNo.3’ does not require
cognitive effort. The process of inspection is explicitly presented as non-symmetrical with
respect to blurred and clear gemstones. Moreover the question worded in frequentist terms
modifies the perception of the situation of revision and sets up the situation of focusing
(Dubois & Prade 1992; Dubois et al. 1996). Participants must simply focus on the reference
class of gemstone on a subset to modify P(gemstone) into P(gemstoneinspection) by
Bayesian conditioning25 . The natural frequencies format improves Participants’ “EBS” rate as
it encourages the comprehension of the focusing situation. In MHPnn-2, if participants envisage
the bi-probabilistic structure, the focusing situation is clearly identifiable in this context where
participants are first asked the frequency question “Has the car appeared behind door No.1 in
numerous repetitions?” and then P(carNo.1’goatNo.3...n’) to the second frequency question:
“Has the car appeared behind door No.1 in numerous repetitions when the host who knows
where the car is, had opened doors No.3; …; No.n to reveal goats? Similarly some other
manipulations may enhance (i) the explicit bi-probabilistic structure and (ii) help participants
to interpret the situation of revision in a focusing framework through a specific wording of the
question. The two theories that have proposed arguments alternative to frequentist module
arguments (the “question form” and the “nested sets”) can be analyzed in this way.
4.2.2 The question form
Girotto & Gonzalez (2001) argue that participants are able to make correct probability
judgment for single-event probabilities when the distributions of probabilities are formulated
in term of chance. Their studies were not concerned with MHPnn-2 so we will stick to the
25 On can noted that Dubois & Prade (1992) use a frequentist framework to explain the focusing situation
33
previous gemstone problem and propose the following statement (see however the studies
cited in Girotto & Gonzalez 2000):
The Gemstone problem (question form): A factory manufactures artificial gemstones daily.
A gemstone that is produced has 2 chances out of 6 to be blurred, 2 chances out of 6 to be
cracked, and 2 chances out of 6 to be neither. It is then tested by the inspection machine. A
blurred gemstone has 1 chance out of 2 to be removed by the inspection machine, a cracked
gemstone has 2 chances out of 2 to be removed by the inspection machine and a clear
gemstone has no chance (out of 2) to be removed by the inspection machine. Imagine that a
gemstone is tested. Out of the total of 6 chances, this gemstone has __ chances of passing the
inspection. What are besides the chances that it has passed the inspection and that it is
blurred?
With this formulation the bi-probabilistic hierarchical structure is again explicit (three
layers in the same way of natural frequencies format and the two distributions of probability
reflected by the given respective proportions). It is also clear that here the “question format”
conducts participants to directly use Laplace’s principle (ratio of proportions) which exactly
corresponds to the focusing process.
4.2.3 Nested set format
Sloman, Over, Slovak and Stibel (2003) propose a general explanation for natural
frequencies effect: framing problems in terms of instances rather than properties actually
clarifies critical set relations. In this line Yamagishi (2003) claims that participants in MHP
tend to answer MR because they cannot comprehend “nested sets” namely, subsets relative to
larger sets, in the task structure. A solution is to propose a schematic representation of
standard MHP (equivalent to Figure 1) (Ichikawa 1989; Cheng & Pitt 2003; Yamagishi
2003). For instance the use of a roulette-wheel diagram (see Figure 3) favors “EBS” in MHP
34
(Ichikawa 1989; Yamagishi 2003). This diagram is a schematic representation of the bi-
probabilistic level of belief described in Figure 1a. The outside circle corresponds to the set of
worlds of layer 0, the inner circle corresponds to the set of worlds of layer 1 and the central
point represents the single set of layer 2. The different areas represent the two distributions of
probabilities related to layers 1 to 0 and to layers 2 to 1. Now if a small ball runs along the
edge of the inner disk and stops at random like a roulette-wheel when participants receive the
message ’goatNo.3’, it is certain that the ball stops in the dotted area, that is, Area carNo.2 or 1/2
of Area carNo.1. This schematic representation, leads participants to correctly interpret
’goatNo.3’ as a focusing message.
-------------------------------------------------------------
Insert Figure 3 about here
-------------------------------------------------------------
4.2.4 Populational format a natural translation of possible worlds semantics
More generally, a populational format of presentation helps participants to induce
experimenters’ representation of MHPnn-2. This format relies on a finite number of discrete
objects (balls) and discrete meta objects (urns) representing worlds and meta-worlds. These
objects (and meta objects) are liable to receive individualizing properties (and meta
properties) on which participants will form beliefs. The traditional hierarchy of balls, urns and
meta-urn (corresponding respectively to worlds of layer 0, layer 1 and layer 2) is the most
convenient introductive set. The focusing situation being here illustrated by the message of
the type “ a ball is extracted of an urn which had been previously extracted, this ball has a
specific characteristic to be named”. The most elementary translation in a populational format
of MHPnn-2 is the urns and balls problem described in section 2.3. Such a representation
indeed facilitates the distinction between the focusing message “the ball is not labeled 3, ... to
35
n” and the updating message “the balls labeled 3, ... to n have been removed” (Baratgin &
Politzer 2007b). The three former formats that improve the interpretation of the focusing
situation can be grouped under this populational representation. For instance Gigerenzer’s
natural frequency format actually corresponds to a general translation (among various others)
of this populational representation.
5 Conclusion
MHPnn-2 appears in the literature as the typical paradigm that concludes that human
probability judgment fails to implement basic revision rule such as Bayesianism conditioning.
However a careful analysis of the experimental statement in the light of the Relevance Theory
and a thorough review of the broad Psychological literature on the subject leads us to
reconsider such conclusion. The heuristics put forward to explain participants’ “non-
Bayesian” response as well as the experimental clues and formats that prompt a higher rate of
Bayesian response can be seen as evidence that participants actually interpret the puzzle in an
updating framework in which the universe undergoes a significant change. Such updating
situation as opposed to the focusing interpretation expected by experimenters requires a
different revision rule which explains the departure from the Bayesian Solution.
MHPnn-2 actually illustrates experimenters’ difficult choice of a methodology in the
studies on probability revision. Before conducting any experimental study, experimenters
should define the consistency principles they retain and outline the implications of the
reference model they use for their empirical work. If an experimenter is to consider the
Bayesian model as the reference to study revision of probabilistic judgment, he or she should
first of all conduct a pragmatic analysis of the representation of the task by participants. More
generally, psychologists should check that participants’ representation of the task coincides
with the type of revision situation they intend (Baratgin & Politzer 2006).
36
In this line MHPnn-2 is the typical example where such a protocol is not respected. First
in the numerous experiments on the puzzle experimenters limit Bayesianism to the simple use
of Bayes’ rule. Participants are judged inconsistent because their responses depart from the
result Bayes’ rule yields on the data given in the statement. Second, experimenters lose sight
aware that Bayes’ rule can be a normative rule in the sole respect of a focusing situation of
revision. Consequently experimenters overlook the fact that participants may not interpret the
puzzle in a focusing framework.
Finally from a psychological point of view, experimenters would advantageously
consider the fairly general Imaging’ rules as a valuable reference to study human probability
judgment. It may actually reveal to have a sizeable descriptive power and robustness as
participants seem to naturally opt for an updating framework. The updating situation of
revision was hardly addressed in the literature. Nevertheless it has been indirectly studied in
three particular research fields. It has been considered explicitly in a deductive framework, in
connection with the problem of belief revision for knowledge bases (Elio & Pelletier 1997;
Politzer & Carles 2001), and implicitly in the field of counterfactual reasoning: a
counterfactual statement indeed expresses a modification of the universe, albeit virtual.
Finally, the situation called “intervention” in the field of causal reasoning can be considered
as a special case of updating (see Appendix 3). Some experimental studies on causal
reasoning have shown that this situation of intervention seems natural to participants as they
are competent to predict the consequences of an intervention (Sloman & Lagnado 2005;
Waldmann & Hagmayer 2005). This new approach to causal reasoning offers additional
reasons to take a serious view at the updating situation in probability judgment revision.
37
Appendix
1 “EBSk” is not unique in general MHPnk
Let MHPnk be the version of the Monty-Hall puzzle with n doors No.i (i = 1, ..., n) and
k the number of doors opened by the host to show goats when participants’ original choice is
No.1 (k ≤ n-2). For parsimony and convenience sake we will order the doors following the
choice of the host and we will assume that the host opens the k last doors (message ‘goatNo.n-
k+1...n’ with k ∈ [1, ..., n-2]). Rules 1, 2, 3, are straightforwardly transposed into MHPnk
while rules 3.2 and 3.2.1 are easily generalized. On the contrary rule 3.1 can have various
generalizations. With respect to rules 3.1, 3.2 and 3.2.1, we assume that the probabilities
P(‘goatNo.n-k+1...n’carNo.i) (for i = 1, ..., i = n) are independent from prior probabilities but
may depend on the remaining number of closed doors (n-k). The likelihoods P(‘goatNo.n-
k+1...n’carNo.i) (i = 1, ..., i = n) are thus function of n and k. Let αi (i = 1, ..., i = n) be the
likelihoods P(‘goatNo.n-k+1...n’carNo.i). Let’s next consider the generalized rules:
Rule 3.2g: If the car is behind door No.1, the host is bound to show goats behind k doors
chosen among doors No.2..., No.n.
Also the rule 3.2.1 (no preference between the n-1 doors if the car is behind door No.1)
becomes:
Rule 3.2.1g: In the specific case where the car is behind door No.1, the host chooses to
show k goats behind doors selected among n-1 doors. The host has thus
€
Cn-1k possibilities and
€
α1 =1Cn-1k .
The general Bayesian solution to MHPnk ,“EBSk” becomes:
38
€
P(car No.i |'goat No.n - k +1...n') = aiα i
a1α1 + α ja jj=2
j=n -k
∑, with 0 < i < n - k +1 (26)
Before going further an assumption needs to be made on the host’s strategies when the car is
behind door No.i (with i>1). We propose first the natural host’s strategy 1.
1.1 Strategy 1: For any i the host adopts the same rule whatever door No.i that hides the
car
In this case the host has no preference between the n-2 doors No.2, ...No.i-1; No.i+1,
...No.n when the car is behind door No.i. Hence the implicit rule 3.1 is no longer ambiguous
and can be generalized by the following rule.
Rule 3.1g1: If the car is behind door No.i (with i > 1), the host chooses to show goats behind k
doors selected among n-2 doors (n doors except doors No.1 and No.i). The host has thus
€
Cn-2k
possibilities and
€
α i =1Cn-2k .
In this case (26) becomes “EBSk1”:
€
P(car No.1 |'goat No.n - k +1...n') =a1
(n −1− k)!k!(n −1)!
a1(n −1− k)!k!
(n −1)!+
(n − 2 − k)!k!(n − 2)!
a jj=2
j=n -k
∑
€
= a1(n −1− k)
a1(n −1− k) + (n −1) a jj=2
j=n -k
∑ (27)
€
P(car No.i |'goat No.n - k +1...n') = ai(n −1)
a1(n −1− k) + (n −1) a jj=2
j=n -k
∑ , for i = 2, ..., i = n (28)
and hence the relation:
€
P(carNo.1 |'goat No.n - k +1...n')P(carNo.i |'goat No.n - k +1...n')
= a1(n −1− k)ai(n −1)
(29)
39
Theorem 8: When the mean of the set of values ai (with i = 2, ...i = n-k) is strictly superior to
the mean of the set of values ai (with i = n-k+1, ..., i = n),
€
P(car No.1 |'goat No.n - k +1...n') < P(car No.1) and in the specific case where the mean of the set
of values ai (with i = 3, ...i = n-k) is equal to the mean of the set of values ai (with i = n-k+1,
...i = n)
€
P(car No.1 |'goat No.n - k +1...n') = P(car No.1) .
Proof. By simple elementary calculation
€
P(car No.1 |'goat No.n - k +1...n') < P(car No.1) ⇔ a1(n −1− k)
a1(n −1− k) + (n −1) a jj=2
j=n -k
∑<
a1
a1 + a jj=2
j=n
∑
€
⇔(n −1)(n −1- k)
−1
aii= 2
i= n -k
∑ > aii= n -k+1
i= n
∑ ⇔
aii= 2
i= n -k
∑(n −1- k)
>
aii= n -k+1
i= n
∑k
In particular when the ai (with i > 1) are equal to a2 for any k, “EBSk2” coincides with “EBS”.
€
P(carNo.1 |'goat No.n - k +1...n') = P(carNo.1 |'goat No.3...n') = P(carNo.1) (30)
Proof. By simply applying relations (27) and (28), we have:
€
P(carNo.1 |'goat No.n - k +1...n') =a1
a1 + (n−1)a2= P(carNo.1) (31)
€
P(car No.i |'goat No.n - k +1...n') = a2(n −1)a1(n −1− k) + a2(n −1- k)(n−1)
= a2(n−1)(n−1− k)(a1 + a2(n −1))
=a2
a1 + a2(n −1)+
ka2
(n −1− k)(a1 + a2(n −1))
€
= P(car No.i) +1
(n − k -1)P(car No.j)
j= n−k +1
n
∑ (32)
The relations (31) and (32) correspond to the solution by simple Imaging rule “SISk”
for MHPnk for the k worlds invalidated by the updating message ‘goatNo.n-k+1...n’ are
redistributed proportionally on the (n-k-1) worlds (carNo.i) (with i = 2, ..., i = n-k).
40
Notably, in standard MHPnk for any k,
€
P(car No.1 |'goat No.n - k +1...n') = P(car No1) = P"'goat No.n-k+1,...,n'
+ (car No.1) (33)
€
for i > 1, P(car No.i |'goat No.n - k +1...n') =(n −1)
(n −1− k)1n
=1n
+k
(n −1− k)n
€
= P(car No.i) +1
(n −1− k)P(car No.j)
j= n−k +1
j= n
∑ = P"'goat No.n-k+1,...,n'
+ (car No.i) (34)
We can again generalize Theorem 4 into
Theorem 9. For any n > 2 and any k,
€
P"goat No.n -k+1...n"(carNo.1) > P(carNo.1 |'goat No.n - k +1...n')
and
€
P'goat No.n -k +1...n' (car No.i) < P(car No.i |'goat No.n - k +1...n'), for i = 2, ...,i = n - k
Proof. “PIS k2” corresponds to
€
P'goat No.n -k +1....n' (car No.i) = ai
a jj=1
n−k
∑, for i = 1, ...,i = n (35)
It is obvious because for n > 2;
€
a1
a jj=1
n−k
∑> a1
a1 +(n−1)
(n−1− k)a j
j= 2
n−k
∑ and ai(n −1)
a1(n −1− k) + (n −1) a jj= 2
n−k
∑>
aj
a jj=1
n−k
∑for i =1, ..., i = n
1.2 Strategy 2: For (any i > 1) the host adopts the same strategy whatever door No.i that
hides the car
This rule is implicit in standard MHP where the host’s strategy is the same for No.2 and
No.3. There are multiple possibilities to generalize the rule of MHP: The host opens door
No.2 if No.3 hides a car and No.3 if No.2 hides a car. A possibility is illustrated below:
Rule 3.1g2: If the car is behind door No.i (with 1 < i < n-k), the host chooses to open door
No.i+1...i+1+k. If the car is behind door No.i (with n-k ≤ i ≤ n), the host chooses to open
41
doors No.i+1...n,2,...2+n-k-i. Hence when the host opens the k doors No.n-k+1, ..., No.n, αn-k
= 1 and αi = 0 otherwise.
In this case (26) becomes “EBSk2”:
€
P(carNo.1 |'goat No.n - k +1...n') =
a1Cn-1k
a1Cn-1k + an-k
= a1a1 + an−kCn-1
k (36)
Similarly
€
P(carNo.n - k |'goat No.n - k +1...n') =an−kCn-1
k
a1 + an−kCn-1k (37)
and hence the relation:
€
P(carNo.1 |'goat No.n - k +1...n')P(carNo.n - k |'goat No.n - k +1...n')
= a1an−kCn-1
k (38)
Theorem 10: (i) For an-k strictly superior to
€
aii= 2, i≠n−k
i= n
∑(Cn -1
k −1),
€
P(car No.1 |'goat No.n - k +1...n') < P(car No.1) = a1
A
and (ii) in the specific case where an-k corresponds to the mean of set of values ai (with i > 2,
and i ≠ n-k) and k = 1 or k = n-2,
€
P(car No.1 |'goat No.n') = P(car No.1 |'goat No.3...n') = P(car No.1) = a1
A since
€
Cn -1k = 2 when k = n − 2 and k =1.
Theorem 11. (i)
€
P(carNo.1 |'goat No.n - k +1...n') < P(carNo.1 |'goat No.n - k +1') =a1
a1 + (n−1)an -k
and (ii) In the special case where all ai are equals, for i > 1, we get from the well know
binomial coefficients property
€
Cn-1k = Cn-1
n -1-k the following result:
For i = 1, ..., i = n:
€
P(carNo.i |'goat No.n - k +1...n') = P(carNo.i |'goat No.k +1...No.n')
42
We have for 1 < k < n-2 the following inequalities:
€
P(carNo.1 |'goat No.n - k +1...n') > P(carNo.1 |'goat No.3...n')
€
P(carNo.n - k |'goat No.n - k +1...n') < P(carNo.n - k |'goat No.3...n')
Proof.
€
a1
a1 + a2Cn -1k > a1
a1 + a2(n −1) because Cn -1
k ≤(n −1)k
k!≤
(n −1)n -2
(n - 2)! < n -1 :
In standard MHPnk, we have:
€
P(carNo.n - k |'goat No.n - k +1...n') =1
1+ Cn-1k (39)
€
P(carNo.n - k |'goat No.n - k +1...n') =Cn-1k
1+ Cn-1k (40)
We can also generalize Theorem 4:
Theorem 12. For any n > 2 and any k,
€
P'goat No.n-k+1,...,n'
+ (carNo.1) > P(carNo.1 |'goat No.n - k +1...n')
and
€
P'goat No.n-k+1,...,n'
+ (carNo.n - k) < P(carNo.n - k |'goat No.n - k +1...n').
Proof. “IPSk2” reads:
€
P'goat No.n-k+1,...,n'
+ (carNo.i) =ai
ai + an-k and it is obvious since for n > 2;
€
a1
a1 + an -k
> a1
a1 + Cn -1k an -k
and Cn -1k an -k
a1 + Cn -1k an -k
>an -k
a1 + an -k
as
€
for k < n -1, Cn -1k ≥
n −1k
k
> 1.
2 Participants’ representation of a distribution of events for MHPk
The possibility though not tested by psychologists, that participants do not have a
precise distribution of probability from layer 1 to layer 0 supports further the hypothesis that
participants interpret the message in an updating manner. In this case, the hierarchical
structure is formed by a probability distributions from layer 2 to layer 1 always represented by
the prior probabilities P(carNo.i) and by a distribution of events from layer 1 to layer 0. This
structure is called a “distribution of events” (Walliser & Zwirn 1997, forthcoming). In this
43
representation the value of event carNo.1 is represented by an interval [Bel(carNo.1),
Pl(carNo.1)] with Bel (carNo.1) the lower value associated to carNo.1 (called “belief function” )
and Pl(carNo.1) is the upper value associate to carNo.1 (called “plausibility”) with relation
Pl(carNo.1) = 1 – Bel(¬carNo.1) . The two measures Bel(carNo.1) and Pl(carNo.1) are defined
with the mass function m(B) that are the weight of evidence in favor of carNo.1.
€
Bel(car No.1) = m(B) = P(car No.1)B: B⊂A≠∅∑ and
€
Pl(car No.1) =1- Bel(¬car No.1) = m(B) =
B:A∧B≠∅∑ P(car No.1).
For such distribution of events the adequate rule of revision after an updating message
‘goatNo.n-k+1...n’ is the Dempster-Shafer’ rule (Dubois & Prade 1993; Walliser & Zwirn
1997, forthcoming) which is expressed by
€
Bel'goat No.n -k+1...n' (carNo.i) =1- Pl'goat No.n -k+1...n' (¬carNo.i)
and
€
Pl'goat No.n -k+1...n' (carNo.i) =Pl(carNo.i∧'goat No.n - k +1...n')
Pl('goat No.n - k +1...n').
Numerous authors have shown that the solution of standard MHP by Dempster-
Shafer’s rule gives again 1/2 for the two measures
€
Bel'goat No.3' (car No.1) and Pl'goat No.3' (car No.1)
(Diagonis & Zabell 1986; Pearl 1988; Fagin & Halpern 1991; Walley 1991; Dubois & Prade
1992). More generally it is immediate to see that the solution of MHPnk by Dempster-Shafer’s
rule (noted “DSk”) is equivalent to “PIS”. Thus we have for the two possibilities of the host’s
strategies treated in Appendix 1 the following “DSk1” and “DSk2”:
€
Bel'goat No.n -k +1...n' (car No.i) = Pl'goat No.n -k +1...n' (car No.i) = ai
aii=1
n -k
∑ , for i =1, ..., i = n (Strategy 1) (41)
€
Bel'goat No.n -k +1...n' (car No.i) = Pl'goat No.n -k +1...n' (car No.i) = ai
ai + an -k
, for i =1, ..., i = n (Strategy 2) (42)
44
As far as focusing situation the solution is not unique. The adequate rule is then the
Fagin Halpern Jaffray’s rule (Dubois & Prade 1993; Walliser & Zwirn 1997, forthcoming)
which defines that:
€
Bel(carNo.i |'goat No.n - k +1...n') =Bel(carNo.i∧'goat No.n - k +1...n')
Bel(carNo.i∧'goat No.n - k +1...n') + Pl(¬carNo.i∧'goat No.n - k +1...n')
€
Pl(carNo.i |'goat No.n - k +1...n') =Pl(carNo.i∧'goat No.n - k +1...n')
Pl(carNo.i∧'goat No.n - k +1...n') + Bel(¬carNo.i∧'goat No.n - k +1...n')
In standard MHP, these rules give a solution in the format of the interval [0, 1/2]
(Diagonis & Zabell 1986; Pearl 1988; Fagin & Halpern 1991; Walley 1991; Dubois & Prade
1992) and
€
0, Pl'goat No.n-k+1...n'
(car No.i)[ ] in MHPnk. Indeed following the host’s strategies, we have
respectively:
€
Pl'goat No.n -k+1...n' (No.i) =aiAi=1
i=n -k
∑
€
= Pl(carNo.i∧'goat No.n - k +1...n') +Bel(carNo.i∧'goat No.n - k +1...n') (Strategy 1) (43)
€
Pl'goat No.n -k+1...n' (No.i) =a1 +an-kA
€
= Pl(carNo.i∧'goat No.n - k +1...n') +Bel(carNo.i∧'goat No.n - k +1...n') (Strategy 2) (44)
3 A correspondence with Pearl’s Bayesian conditioning after intervention
The proportional Imaging rule is equivalent to Pearl (2000)’ s Bayesian conditioning
after one intervention (designed by Pearl as the “do” operator) introduced in the causal
reasoning. MHPnn-2 can indeed be modeled by a Bayesian network representing dependencies
among: the host’s initial choice (X1), participants’ choice (X2) and the doors open by host
45
hiding goats (X3) (see Figure 4a below). All variables X1, X2 and X3 can take n values: door
No.1, ..., door No.n.
Now if the external intervention “the n-2 doors No.3, ..., No.n are open” is released
(do(X3) = No.3...n), the links X1 → X3 and X1→ X3 are deleted and the new mechanism
(when X2 is equal to No.1) corresponds to conditioning on the action do(X3) = No.3...n (see
Figure 4b). We find the relations (9) and (10) induced by the proportional Imaging rule:
€
P(X1=No.1 | (X2 =No.1)∧do(X3) = No.3...n) = P(carNo.1|goat No.3...n)
€
= P'goat No.3,...,n'
+ (car No.1) (43)
€
P(X1=No.2 | (X1 =No.1)∧do(X3) = No.3...n) = P(carNo.2|goat No.3...n)
€
= P'goat No.3,...,n'
+ (car No.2) (44)
-------------------------------------------------------------
Insert Figure 4 about here
-------------------------------------------------------------
This correspondence supports a new psychological argument in favor of the
hypothesis of updating interpretation by participants. It explains the hypothesis introduced by
Glymour (2001) and studied by Burns and Wieth (2004) that participants are not able to
recognize the collider principle. The “collider principle” is depicted in Figure 4a. The two
variables X1 (experimenter’s original choice) and X2 (participant’s original choice) are
independent whereas the value of variable X3 (doors hiding goats opened by the host) renders
X1 and X2 dependent. Now, in the updating representation of MHPnn-2 (Figure 4b), the
collider principle is ruled out. Hence participants if have an updating interpretation cannot be
detected the collider principle.
46
Acknowledgements
I thank G. Politzer, B. Walliser and two anonymous referees for their fruitful comments.
47
References
Baratgin, J., 1999. Psychologie du raisonnement probabiliste dynamique. Recherches
historiques, pragmatiques et expérimentales sur la règle de Bayes. [The psychology of
dynamic probabilistic reasoning. Historical, pragmatic, and experimental research on
Bayes’ rule]. Doctorate dissertation. Ecole Polytechnique.
Baratgin, J., 2004. Redundancy effect in the revision of degrees of belief. In: ICT_2004. The
Fifth International Conference on Thinking, Louvain, Belgique
Baratgin, J., Politzer, G., 2003. Monty hall puzzle: A pragmatic interpretation of participants'
usual revision response. In: Escop-2003, Grenada, Spain
Baratgin, J., Politzer, G., 2006. Is the mind Bayesian? The case for agnosticism. Mind &
Society 5, 1-38
Baratgin, J., Politzer, G., 2007a. The Psychology of Dynamic Probability Judgment: Order
Effect, Normative Theories, and Experimental Methodology. Mind & Society 6, 53-66
Baratgin, J., Politzer, G., 2007b. Updating: A psychologically basic situation of probability
revision. MS, Institut Jean Nicod Paris
Barbey, A.K., Sloman, S., Forthcoming. Base-rate respect: From ecological rationality to dual
processes Behavioral and Brain Sciences
Bertrand, J., 1889. Calcul des probabilités [Theory of probability]. Paris.
Billot, A., Walliser, B., 1999. Epistemic properties of knowledge hierarchies. Journal of
Mathematical Economics 32, 185-205
Brase, G.L., 2002. Ecological and evolutionary validity: Comments on Johnson-Laird,
Legrenzi, Girotto, Legrenzi, and Caverni’s (1999) Mental-Model Theory of
Extensional Reasoning Uncertainty. Psychological Review 109, 722-728
Brase, G.L., Cosmides, L., Tooby, J., 1998. Individuation, counting and statistical inference:
the role of frequency and whole-object representations in judgment under uncertainty.
journal of Experimental Psychology: General 127, 3-21
Burns, B.D., Wieth, M., 2004. The collider principle in causal reasoning: Why the Monty
Hall dilemma is so hard. Journal of Experimental Psychology: Learning, Memory and
Cognition 133, 434-449
Caplan, B., 2000. Rational irrationality: A framework for the neoclassical-behavioral debate.
Eastern Economic Journal 26, 191-211
48
Cheng, P.C.H., Pitt, N.G., 2003. Diagrams for difficult problems in probability. Mathematical
Gazette 87, 86-97
Chun, Y.H., 2003. On the information economic approach to the generalized show problem.
American Statistical Association 53, 43-51
Cosmides, L., Tooby, J., 1996. Are humans good intuitive statisticians after all? Rethinking
some conclusions from the literature on judgment under uncertainty. Cognition 58, 1-
73.
Cross, C.B., 2000. A characterization on imaging in terms of Popper functions. Philosophy of
Science 67, 316-338
de Finetti, B., 1949. Rôle et domaine d'application du théorème de Bayes selon les différents
points de vue sur les probabilités. In: Actes du XVIIIe colloque de calcul des
probabilités, pp. 67-83, Paris
de Finetti, B., 1972. Probability, Induction and Statistics. The art of guessing. Wiley, New-
York.
Diagonis, P., Zabell, S., 1986. Some alternatives to Bayes' rule. In: Grofman B & Owen G
(eds.) Information and group decision making. Jai Press, Greenwich, CT, pp. 25-38.
Dubois, D., Prade, H., 1992. Evidence, knowledge, and belief function. International Journal
of Approximate Reasoning 6, 295-319
Dubois, D., Prade, H., 1993. Belief revision and updates in numerical formalisms. In:
Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence
(IJCAI-93), pp. 620-625, Chambéry, France
Dubois, D., Prade, H., 1997. Focusing vs. belief revision: A fundamental distinction when
dealing with generic knowledge. In: Lecture Notes In Computer Science. Springer-
Verlag, London, UK, pp. 96-107.
Dubois, D., Prade, H., Smets, P., 1996. Representing partial ignorance. IEEE System Machine
and Cybernetic 26, 361-377.
Ducrot, O., 1972. Dire et ne pas dire [To tell and not to tell]. Hermann, Paris.
Dulany, D.L., Hilton, D.J., 1988. Conversational implicature, conscious representation, and
the conjunction fallacy. Social Cognition 9, 85-100
Elio, R., Pelletier, F.J., 1997. Belief change as propositional update Cognitve Science 21,
419-460
Fagin, R., Geanakoplos, J., Halpern, J.Y., Vardi, M.Y., 1999. The hierarchical approach to
modeling knowledge and common knowledge. International Journal of Game Theory
28, 331-365
49
Fagin, R., Halpern, J.Y., 1991. A new approach to updating beliefs. In: Bonissone M, Henrion
LN, Kanal JF & Lemmer J (eds.) Uncertainty in Artificial Intelligence. Elsevier
Science Publishers, New York, NY, pp. 347-374.
Falk, R., 1992. A closer look at the probabilities of the notorious three prisoners. Cognition
43, 197-223
Fillenbaum, S., 1976. Inducements: On the phrasing and logic of conditional premises,
threats, and warnings. Psychological Research 38, 231-250
Fox, C.R., Clemen, R.T., 2005. Subjective probability assessment in decision analysis:
Partition dependence and bias yoward the ignorance prior. Management Science 51,
1417-1432
Fox, C.R., Levay, J., 2004. Partition edit count: naive extensional reasoning in judgment of
conditional probability. Journal of Experimental Pychology: General 133, 626-642
Fox, C.R., Rottenstreich, Y., 2003. Partition priming in judgment under uncertainty.
Psychological Science 14, 195-200
Franco-Watkins, A.M., Derks, P.L., Dougherty, M.R.P., 2003. Reasoning in the Monty Hall
problem: Examining choice behavior and probability judgment. Thinking &
Reasoning 9, 67-90
Friedman, D., 1998. Monty Hall's three doors: Construction and deconstruction of a choice
anomaly. The American Economic Review 88, 933-946
Gärdenfors, P., 1988. Knowledge in Flux, Modeling the Dynamics of Epistemic States. The
MIT Press., Cambridge, MA.
Gardner, M., 1959. Mathematical games: How three mathematicians disproved a celebrated
conjecture of Leonard Euler. Scientific American 201, 181-188
Geis, M.L., Zwicky, A.M., 1971. On invited inferences. Linguistic Inquiry 2, 561-566
Gigerenzer, G., Hoffrage, U., 1995. How to improve Bayesian reasoning without instruction:
Frequency formats. Psychological Review 102, 684-704
Gigerenzer, G., Selten, R., 2001. Bounded rationality: The adaptive toolbox.
Gillman, L., 1992. The car and the goats. American Mathematical Monthly 99, 3-7
Gilovich, T., Griffin, D.W., Kahneman, D., 2002. Heuristic and biases: The psychology of
intuitive judgment. Cambridge University Press, New York.
Gilovich, T., Medvec, V.H., Chen, S., 1995. Commission, omission, and dissonance
reduction: Coping with regret in the "Monty Hall" problem. Personality and Social
Psychology Bulletin 21, 182-190
50
Girotto, V., Gonzalez, M., 2000. Strategies and models in statistical reasoning. In: Schaeken
W, De Vooght G, Vandierendonck A & d'Ydewalle G (eds.) Deductive reasoning and
strategies. Mahwah, N.J.: Lawrence Erlbaum Associates, pp. 267-285.
Girotto, V., Gonzalez, M., 2001. Solving probabilistic and statistical problems: A matter of
information structure and question form. Cognition 78, 247–276
Glymour, C., 2001. The mind's arrows: Bayes nets and graphical causal models in
psychology. The PIT Press, Cambridge.
Granberg, D., 1996. A new version of the Monty Hall Dilemma with unequal probabilities.
Behavioral Processes 48, 35-34
Granberg, D., Brown, T.A., 1995. The Monty Hall dilemma. Personality and Social
Psychology Bulletin 21, 711-723
Grünwald, P.D., Halpern, J.Y., 2003. Updating probabilities. Journal of Artificial Intelligence
Research 19, 243-278
Hammerton, M., 1973. A case of radical probability estimation. Journal of Experimental
Psychology: General 101, 252-254
Hoffrage, U., Gigerenzer, G., Krauss, S., Martignon, L., 2002. Representation facilitates
reasoning: What natural frequencies are and what they are not. Cognition 84, 343-352
Ichikawa, S., 1989. The role of isomorphic schematic representation in the comprehension of
counterintuitive Bayesians problems. Journal of Mathematical Behavior 8, 269-281
Ichikawa, S., Takeichi, H., 1990. Erroneous belief in estimating posterior probability.
Behaviormetrika 27, 59-73
Jeffrey, R., 1988. Conditioning, kinematics and exchangeability. In: Skyrms B & Harper W
(eds.) Causation, chance and credence. Reidel, pp. 221-255.
Johnson-Laird, P.N., Legrenzi, P., Girotto, V., Sonino-Legrenzi, M.S., 1999. Naive
probability: A mental model theory of extensional reasoning. Psychological Review
106, 62-88
Kahneman, D., Tversky, A., 1972. On prediction and judgment. ORI (Oregon research
Institute) 12
Kahneman, D., Tversky, A., 1973. On the psychology of prediction. Psychological Review 80
Katsuno, A., Mendelzon, A., 1992. On the difference between updating a knowledge base and
revising it. In: Gärdenfors P (ed.) Belief Revision. Cambridge University Press,
Cambridge, pp. 183-203.
51
Kluger, B.D., Wyat, S.B., 2004. Are judgment errors reflected in market prices and
allocations, Experimental evidence based in the Monty Hall problem. The Journal of
Finance 3, 969-997
Krauss, S., Wang, X.T., 2003. The Psychology of the Monty Hall problem: Discovering
psychological mechanisms for solving a tenacious brain teaser. Journal of
Experimental Psychology: General 132, 3-22
Krosnick, J.A., Li, F., Lehman, D.R., 1990. Conversational conventions, order of information
acquisition, and the effect of base rates and individuating information on social
judgments. Journal of Personality and Social Psychology 39, 1140-1152
Laplace, P.S., 1986. Essai philosophique sur les probabilités [Philosophical essay on the
probabilities]. Bourgeois, C., Paris.
Lepage, F., 1997. Conditionals, Imaging, and subjunctive probability. Dialogue XXXVI, 113-
135
Lewis, D., 1976. Probabilities of conditionals and conditional probabilities. Philosophical
Review 3, 297-315
Loschi, R.H., Iglesias, P.L., Wechsler, S., 2006. Unpredictability, probability updating and
the three prisoners paradox. In: ICOTS 7 Rossman A & Chance B (Eds.). IASE,
Salvador, Bahia, Brazil
Macchi, L., 1995. Pragmatic aspects of the base-rate fallacy. Quarterly Journal of
Experimental Psychology 48A, 188-207
Nickerson, R.S., 1996. Ambiguities and unstated assumptions in probabilistic reasoning.
Psychological Bulletin 120, 410-433
Pearl, J., 1988. Probabilistic reasoning in intelligent system: Networks of Plausible Inference.
Morgan Kaufmann, San Mateo.
Pearl, J., 2000. Causality: Models, reasoning, and inference. Cambridge University Press,
Cambridge.
Piattelli-Palmerini, M., 1994. Inevitable illusions. Wiley, New-York.
Politzer, G., Carles, L., 2001. Belief Revision and Uncertain Reasoning. Thinking and
Reasoning 7, 217-234
Politzer, G., Macchi, L., 2005. The representation of the task: The case of the Lawyer-
Engineer problem. In: Girotto V & Johnson-Laird PN (eds.) The shape of reason.
Essays in honour of P. Legrenzi. Psychology Press, Hove, pp. 119-135.
Politzer, G., Noveck, I., 1991. Are conjunction rule violations the result of conversational rule
violations? Journal of Psycholinguistic Research 20, 83-103
52
Risinger, D.M., Loop, J.L., 2002. Three card monte, Monty Hall, modus operandi and
“offender profiling”: Some lessons of modern cognitive science for the law of
evidence. Cardozo law review 24, 193-285
Shimojo, S., Ichikawa, S., 1989. Intuitive reasoning about probability: Theoretical and
experimental analyses of the problem of three prisoners. Cognition 32, 1-24
Sloman, S.A., Lagnado, D.A., 2005. Do we "do"? Cognitive Science 29, 5–39
Sloman, S.A., Over, D., Slovak, L., Stibel, J.M., 2003. Frequency illusions and other fallacies.
Organizational Behavior and Human Decision Processes 91, 296-309
Sperber, D., Wilson, D., 1995. Relevance: Communication and cognition. Blackwell, Oxford.
Tubau, E., Alonso, D., 2003. Overcoming illusory inferences in a probabilistic
counterintuitive problem: The role of explicit representations. Memory & Cognition
31, 596-607
Waldmann, M., Hagmayer, Y., 2005. Seeing versus doing: Two modes of accessing causal
knowledge Journal of Experimental Psychology: Learning, Memory, and Cognition
31, 216-227
Walley, P., 1991. Statistical reasoning with imprecise probabilities. Chapman and Hall,
London.
Walliser, B., Zwirn, D., 1997. Les règles de révision des croyances [The revision rules of
beliefs]. In: Dupuy JP & Livet P (eds.) Les limites de la rationalité [The limits of
rationality]. La découverte, Paris, pp. 190-208.
Walliser, B., Zwirn, D., 2002. Can Bayes’ rule be justified by cognitive rationality principles?
Theory and Decision 53, 95-135
Walliser, B., Zwirn, D., forthcoming. Change rules for hierarchical beliefs. International
Journal of Approximate Reasoning
Wechsler, S., Esteve, L.G., Simonis, A., Peixoto, C., 2005. Indifference, neutrality and
informativeness: Generalizing the three prisoners paradox. Synthese 143, 255-272
Yamagishi, K., 2003. Faciliting normative judgments of conditional probability: Frequency or
nested sets? Experimental Psychology 50, 97-106
53
54
55
56
top related